Multi-scale dilated convolution of convolutional neural network for image denoising

Convolutional Neural Network has achieved great success in image denoising. The conventional methods usually sense those beyond scope contextual info at the expense of the receptive filed shrinking, which easily lead to multiple limitations. In this paper, we have proposed a concise and efficient convolutional neural network naming Multi-scale Dilated Convolution of Convolutional Neural Network (MsDC), which attempt to utilize the newly designed multi-scale dilated convolution strategy to handle the above mentioned obstinate limitation. The proposed multi-scale dilated convolution module uses the dilated filters to systematically aggregate multi-scale contextual information without reducing the receptive field. The behind rationale of our method is based on the phenomenon that the dilated convolution can effectively expand the corresponding receptive field while conserving those valuable contextual information. Meanwhile, we also utilize residual learning method to learn the residuals directly to speed up the learning procedur. Compared to the state-of-the-art methods, the results have suggested that our method can remove image noise more effectively and efficiently. Our MsDC code can be download at https://github.com/doctorwgd/MsDC.